Qwen3-8B TMax AENV v39b — iteration 189

Reinforcement-learning checkpoint from the tmax_aenv_v39b run, taken at training iteration 189.

The policy was initialized from willamazon1/qwen3-8b-tmax-sft-v3-iter353 (itself an SFT of Qwen/Qwen3-8B) and trained with group-relative policy optimization on asynchronous multi-turn agentic-environment rollouts.

Checkpoints are saved every 10 iterations; iterations 69–219 of this run are published as separate repos in the same collection, so you can compare points along the training curve.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "willamazon1/qwen3-8b-tmax-aenv-v39b-iter189"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16, device_map="auto")

msgs = [{"role": "user", "content": "What is 12*8?"}]
text = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
ids = tok(text, return_tensors="pt").input_ids.to(model.device)
out = model.generate(ids, max_new_tokens=256)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

Conversion

Converted from a Megatron-LM torch_dist training checkpoint to HuggingFace safetensors with slime's tools/convert_torch_dist_to_hf.py. Embedding padding was stripped back to the tokenizer's vocab_size of 151936, so the tensor shapes match upstream Qwen/Qwen3-8B exactly. Weights are bfloat16; the optimizer state of the training checkpoint is not included.

Every shard was checked for NaN/Inf (none found) and the model was loaded and sampled from before upload.